• This section solves the original 30 KB patch route without buying endlessly larger machines. 1:51:53 1:52:19
  • The new c8gn.48xlarge "beast" instances have 192 CPU cores, 384 GB RAM, and 600 Gbit/s network performance, at about $11/hour or $8,000/month each. 1:52:41 1:53:02 1:53:25
  • The Node test runs about 180 processes instead of 192 to leave some cores free. 1:54:02
  • With Node/Cpeak on the beast server, the route moves more than 20 GB/s, about 160 Gbit/s, but CPU becomes the limiting factor before the 600 Gbit/s network is saturated. 1:55:15 1:55:35
  • Express on the same heavier route can barely reach about half a million requests per second. 1:56:30 1:56:53
  • Cpeak or Fastify can get close, but Node's parent-process distribution overhead blocks the final jump to 1 million per second. 1:57:24 1:56:13
  • The route is rewritten in C++ with Drogon, one of the fastest web frameworks the uploader knows. 1:58:05 1:58:24
  • Drogon's default JSON parser is initially slower than Node/V8 for this case, so the uploader switches to RapidJSON. 1:58:56 2:00:04
  • Compression and logging are disabled to keep the test comparable and preserve the intended 30 KB response size. 2:00:34 2:00:52

Result

  • The C++ server averages about 1 million requests per second over 60 seconds and reaches about 1.2 million in better samples. 2:03:27
  • The test moves about 38 GB/s, roughly 300 Gbit/s, and transfers about 2 TB in one minute at the application layer. 2:03:54 2:05:05
  • The C++ process uses far less memory than the Node setup and leaves about 30% CPU idle, suggesting another limit remains. 2:03:16 2:06:12
  • The uploader estimates C++ gives about a 30% request-per-second boost over raw Node/Cpeak while using less CPU. 2:06:30
  • Attempts with load balancing across two beast machines do not improve results because the load balancer or system/network limits become the next bottleneck. 2:07:51 2:08:18